Fast Online Video Pose Estimation by Dynamic Bayesian Modeling of Mode Transitions

Area of Science:

• Computer Vision
• Machine Learning
• Human Pose Estimation

Background:

• Human pose estimation from videos is crucial for many applications.
• Existing methods often struggle with real-time processing or lack robustness.

Purpose of the Study:

• To develop a fast, online method for detecting and tracking human upper-body poses in live video streams.
• To improve accuracy and efficiency in video pose estimation, especially in high frame rate scenarios.

Main Methods:

• Utilizes a time-windowed conditional dynamic Bayesian network (CDBN) for pose inference.
• Models latent pose modes and their transitions using current observations, probabilistic priors, and multimode softmax regression.
• Leverages temporal correlations between video frames for robust estimation.

Main Results:

• Achieves improved accuracy and efficiency compared to existing online video pose estimation methods.
• Demonstrates effectiveness on multiple benchmark datasets, including a new high frame rate dataset.
• Successfully captures action-related temporal information in high frame rate scenarios.

Conclusions:

• The proposed time-windowed CDBN method offers a robust and efficient solution for online human pose estimation.
• The approach is well-suited for real-time applications and high frame rate video analysis.
• This work advances the state-of-the-art in video-based human pose tracking.